Application of Multiple Neural Networks in Aspect Sentiment Analysis

Abstract

Sentiment Analysis (SA) has been an important part of social media data analysis; it provides more accurate information and classification performance. Common methods of sentiment analysis use machine learning or text information classification based on sentiment dictionary, but these methods do not effectively capture the subject of a sentence. Usually there are some keywords which called ‘target’ are considered important emotional information in a sentence. The target words and its corresponding emotional vocabulary reflect the emotional situation of a sentence in different aspects, so we call it Aspect-Based Sentiment Analysis (ABSA). ABSA aims to predict fine-grained emotions in different aspects of comments or reviews. It is no longer a simple judgment of the emotional tendency of a sentence, good or bad, but a deeper analysis of the different aspects of emotion mentioned. Neural networks have also been widely used in sentiment analysis at the level of aspects. To explore it, we propose two methods to verify it. First, we try to use Bi-directional Long Short-Term Memory network (Bi-LSTM) to replace the existing Long Short-Term Memory networks (LSTM). The second method is to try to combine Convolutional Neural Network (CNN) with LSTM. We apply both methods to document-level data for pre-training to obtain parameters, through transfer learning we apply parameters to formal training. The experiments results show multiple neural network perform well in aspect-based sentiment analysis.